Scholarlink

<p>Object detection in drone imagery presents substantial challenges in balancing high accuracy with computational efficiency for on-board deployment. To address this, we propose EBAD-YOLO (Efficient Bidirectional Adaptive Dense YOLO), a lightweight architecture built upon YOLOv10s. Our design introduces a Bidirectional Adaptive Dense Connection Feature Pyramid Network (BADC-FPN) that enables effective multi-scale feature fusion through bidirectional cross-scale dense connections and adaptive weighting. To enhance computational efficiency, we optimize the C2f module by integrating the lightweight FasterBlock. Additionally, a Distribution-Guided Quality Predictor (DGQP) is incorporated into the detection head to suppress low-quality predictions and improve the precision–recall trade-off. Finally, layer-adaptive magnitude-based pruning (LAMP) is applied to further compress the model for efficient edge deployment. On the VisDrone2019 dataset, EBAD-YOLO achieves an mAP@50 of 35.9%, a 3.6 percentage-point improvement over YOLOv10s, while reducing GFLOPs by 56.5% and parameters by 66.7%. On edge devices, it delivers 88.3&#xa0;FPS on the NVIDIA Jetson Orin NX and 22.3&#xa0;FPS on the Jetson Nano, enabling real-time aerial detection under stringent power and memory constraints. Its strong generalization capability is further validated on the TinyPerson and LEVIR-Ship datasets.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

EBAD-YOLO: efficient bidirectional adaptive dense network for UAV small-object detection

  • 期刊论文

  • 发表日期：

    2025年12月26日

查看全文
  • Shuo Hu,
  • Run Xing,
  • Liyang Han,
  • Tongtong Liu

摘要

Object detection in drone imagery presents substantial challenges in balancing high accuracy with computational efficiency for on-board deployment. To address this, we propose EBAD-YOLO (Efficient Bidirectional Adaptive Dense YOLO), a lightweight architecture built upon YOLOv10s. Our design introduces a Bidirectional Adaptive Dense Connection Feature Pyramid Network (BADC-FPN) that enables effective multi-scale feature fusion through bidirectional cross-scale dense connections and adaptive weighting. To enhance computational efficiency, we optimize the C2f module by integrating the lightweight FasterBlock. Additionally, a Distribution-Guided Quality Predictor (DGQP) is incorporated into the detection head to suppress low-quality predictions and improve the precision–recall trade-off. Finally, layer-adaptive magnitude-based pruning (LAMP) is applied to further compress the model for efficient edge deployment. On the VisDrone2019 dataset, EBAD-YOLO achieves an mAP@50 of 35.9%, a 3.6 percentage-point improvement over YOLOv10s, while reducing GFLOPs by 56.5% and parameters by 66.7%. On edge devices, it delivers 88.3 FPS on the NVIDIA Jetson Orin NX and 22.3 FPS on the Jetson Nano, enabling real-time aerial detection under stringent power and memory constraints. Its strong generalization capability is further validated on the TinyPerson and LEVIR-Ship datasets.